The Geometry of Trust: Verifiable Value Alignment via Causal Manifold Structure

Read full text Summarise with VIRAT
Authors: Jade Wilson
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We introduce the causal inner product ⟨u, v⟩_c = uᵀ(UᵀU)v, derived from a transformer's unembedding matrix, as a metric for inspecting the internal geometry of value representations in large language models. Empirical measurements across eight models (GPT-2, GPT-2-medium, Qwen 2.5-0.5B, TinyLlama-1.1B, StableLM-3B, and their alignment-tuned variants) show that value-relevant geometric structure — semantically meaningful clustering, curvature distinguishing affective from structural values, and non-trivial manifold dimensionality — is absorbed during pretraining and largely invariant to post-hoc alignment training including SFT, DPO, and PPO-based RLHF. Control experiments against matched non-value term sets confirm this structure is specific to value-laden content. An open-source Rust proof-of-concept (https://github.com/jade-codes/got) demonstrates that these measurements are computable, deterministic, and independently reproducible.

Fields of Study & Subjects

Topic Modeling Explainable Artificial Intelligence (XAI) Ethics and Social Impacts of AI

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
EN